Integrated: 8352587: C2 SuperWord: we must avoid Multiversioning for PeelMainPost loops

[Emanuel Peter]

On Mon, 24 Mar 2025 08:22:57 GMT, Emanuel Peter <epeter at openjdk.org> wrote:

> This was a fuzzer failure, which hit an assert in SuperWord:
> 
> `# assert(_cl->is_multiversion_fast_loop() == (_multiversioning_fast_proj != nullptr)) failed: must find the multiversion selector IFF loop is a multiversion fast loop`
> 
> We had a fast main loop, but it could not find the `multiversion_if`. The reason was that the loop was a `PeelMainPost` loop, i.e. there is no pre-loop but only a single peeled iteration. This makes the pattern matching from main-loop via pre-loop to `multiversion_if` impossible.
> 
> I'm proposing two changes in this PR:
> - We must check `peel_only`, to see if we are in a `PeelMainPost` or `PreMainPost` case, and only do multiversioning if we know that there will be a pre-loop.
> - In `eliminate_useless_multiversion_if` we should already detect that a main-loop that is marked as multiversioned should be able to find its `multiversion_if`. I'm removing its multiversioning marking if we cannot find the `multiversion_if`.
> 
> I added 2 tests:
> - The fuzzer generated test that hits the assert before this patch.
> - An IR test that checks that we do not multiversion in a  `PeelMainPost` loop case.
> 
> ---------------
> 
> **FYI**: I tried to add an assert in `eliminate_useless_multiversion_if` that we must always find the `multiversion_if` from a multiversioned main loop. But there are cases where this can fail. Here an example:
> 
> `test/hotspot/jtreg/compiler/locks/TestSynchronizeWithEmptyBlock.java`
> 
> With flags: `-ea -esa -XX:CompileThreshold=100 -XX:+UnlockExperimentalVMOptions -server -XX:-TieredCompilation`
> 
> 
> Counted            Loop: N537/N176  counted [int,100),+1 (-1 iters) 
> Loop: N0/N0  has_sfpt
>   Loop: N307/N361  limit_check profile_predicated predicated sfpts={ 182 495 }
>     Loop: N536/N535 
>       Loop: N537/N176  counted [int,100),+1 (-1 iters)  has_sfpt strip_mined
>     Loop: N379/N383  limit_check profile_predicated predicated counted [int,int),+1 (4 iters)  pre rc  has_sfpt
>     Loop: N353/N357  counted [int,1000),+1 (4 iters)  post rc  has_sfpt
> Multiversion       Loop: N537/N176  counted [int,100),+1 (100 iters)  has_sfpt strip_mined
> PreMainPost        Loop: N537/N176  counted [int,100),+1 (100 iters)  multiversion_fast has_sfpt strip_mined
> Unroll 2           Loop: N537/N176  counted [int,100),+1 (100 iters)  main multiversion_fast has_sfpt strip_mined
> Poor node estimate: 306 >> 92
> Loop: N0/N0  has_sfpt
>   Loop: N307/N361  limit_check profile_predicated predicated sfpts={ 182 }
>     Loop: N556/N557  sfpts={ 559 }
>       Loop: N552/N554  counted...

This pull request has now been integrated.

Changeset: c856b342
Author:    Emanuel Peter <epeter at openjdk.org>
URL:       https://git.openjdk.org/jdk/commit/c856b3425a70d2aecb6c5c44da36396a5d74b00d
Stats:     139 lines in 5 files changed: 134 ins; 0 del; 5 mod

8352587: C2 SuperWord: we must avoid Multiversioning for PeelMainPost loops

Reviewed-by: chagedorn, kvn

-------------

PR: https://git.openjdk.org/jdk/pull/24183